Transform class model coding method and system for channel state information prediction

Abstract

The invention provides a coding method and system of a Transform class model for channel state information prediction. The coding method comprises the following steps: S1, acquiring and processing channel state time sequence data; s2, inputting the processed data into an encoder to extract features; s3, CFR data needing to be predicted are embedded, and position codes of the model are added; decoding the relevant parameters using a decoder; s4, processing the output of the decoder, and predicting channel state information; and S5, taking the CFR data and the CSI data obtained in the actual scene as the input of the model, and predicting channel frequency response information. According to the method, the time sequence characteristics of the channel state information are calculated and coded, and the time information is brought into the model, so that the relative stability time of the channel at the future moment is considered, the problem of weak time sequence information embedding capability in a Transform model is solved, and the accuracy of channel state prediction is improved.

Description

Technical Field

[0001] The present invention relates to the field of channel state information prediction, and in particular to a coding method and system for a Transformer-type model for channel state information prediction. Background Art

[0002] Accurately predicting Channel State Information (CSI) has become a key factor in improving network performance and enabling effective radio resource management. Channel prediction is crucial for link adaptation, as it minimizes capacity loss caused by channel variations and enables more efficient utilization of available bandwidth.

[0003] In recent years, deep learning-based methods have brought new possibilities to the task of channel state time series prediction. Researchers have attempted to use models such as RNN, CNN, and LSTM to predict channel states. Among these models, Transformer-based models have become a popular solution.

[0004] However, Transformer-based models all use attention structures, which makes them unable to capture temporal order information. Therefore, Transformer-based models require position encoding to help the model obtain positional order information. Ordinary position encoding has limited use of temporal information and can only mark the position information of input features.

[0005] A Chinese patent publication with publication number CN117674926A discloses a method and apparatus for processing channel state information. The method comprises: inputting channel state information into an encoder model to obtain extended and enhanced channel state information output by the encoder model; and transmitting the extended and enhanced channel state information to a network device. The encoder model comprises an encoder and a channel state information extension and enhancement module cascaded with the encoder. This patent document aims to enable reuse of AI models under different parameter configurations, reduce the complexity of AI model training, and conserve AI model deployment resources, but it still fails to address the aforementioned issues. Summary of the Invention

[0006] In view of the defects in the prior art, the object of the present invention is to provide a coding method and system for a Transformer-type model for channel state information prediction.

[0007] According to the present invention, a method for encoding a Transformer-type model for channel state information prediction includes:

[0008] Step S1: Acquire and process channel state time series data;

[0009] The time series data includes CSI data and CFR data; the processing includes embedding and encoding;

[0010] Step S2: Input the processed data into the encoder to extract features;

[0011] Step S3: Embed the CFR data to be predicted and add the position encoding of the Transformer model; input the relevant parameters into the decoder for decoding;

[0012] The parameters include the output of the encoder and the CFR feature data to be predicted;

[0013] Step S4: Process the output of the decoder to predict channel state information;

[0014] Step S5: Use the CFR data and CSI data obtained in the actual scenario as input to the model to predict channel frequency response information.

[0015] Preferably, the step S1 includes the following sub-steps:

[0016] Step S1.1: Split the CFR data into real and imaginary parts and concatenate them into a real matrix:

[0017]

[0018] in, is the real part of the CFR data, is the imaginary part of the CFR data, concat(·) is the concatenation function;

[0019] Step S1.2: Calculate and obtain the time code;

[0020] Step S1.3: The CFR data is embedded through the convolutional layer, and the time encoding and position encoding of the Transformer model are added.

[0021] Preferably, step S4 includes the following sub-steps:

[0022] Step S4.1: Input the decoder output into the fully connected layer, and crop the fully connected layer output to obtain data with the same dimension as the prediction as the prediction result;

[0023] Step S4.2: Use the mean square error function as the loss function, calculate the error between the predicted value and the true value and perform backpropagation to train the network.

[0024] Preferably, the step S1 includes using CFR data as a signal for a channel state prediction task and using CSI data to calculate time coding.

[0025] Preferably, the step S1.2 includes the following sub-steps:

[0026] Step S1.2.1: Calculate the differential L1 and L2 norms of the CSI signal CSI∈C t×n , the calculated L1 and L2 norms CSIL∈R t :

[0027]

[0028]

[0029] Where t is the length of the time series, n is the frequency, abs(·) refers to the amplitude of the complex number, and i and j are array position indices;

[0030] Step S1.2.2: The rate of change η of the time series position t relative to the position p t,p for:

[0031]

[0032] Step S1.2.3: Use the rate of change η t,p And the selected threshold η0 is calculated to obtain the TC time code:

[0033]

[0034] Among them, i and j are array position indexes;

[0035] Step S1.2.4: Use a linear layer to map the temporal code TC to the same dimension as the embedded CFR feature. After mapping, TC∈R t×d ;

[0036] Among them, d is the dimension after mapping.

[0037] According to the present invention, a coding system of a Transformer-type model for channel state information prediction is provided, comprising:

[0038] Module M1: Acquire and process channel state timing data;

[0039] The time series data includes CSI data and CFR data; the processing includes embedding and encoding;

[0040] Module M2: Inputs the processed data into the encoder to extract features;

[0041] Module M3: Embeds the CFR data to be predicted and adds the position encoding of the Transformer model; inputs the relevant parameters into the decoder for decoding;

[0042] The parameters include the output of the encoder and the CFR feature data to be predicted;

[0043] Module M4: processes the decoder output and predicts channel state information;

[0044] Module M5: Uses the CFR data and CSI data obtained in actual scenarios as input to the model to predict channel frequency response information.

[0045] Preferably, the module M1 includes the following submodules:

[0046] Module M1.1: Split the CFR data into real and imaginary parts and concatenate them into a real matrix:

[0047]

[0048] in, is the real part of the CFR data, is the imaginary part of the CFR data, concat(·) is the concatenation function;

[0049] Module M1.2: Calculate time code;

[0050] Module M1.3: Embed the CFR data through the convolutional layer, add time encoding and position encoding of the Transformer model.

[0051] Preferably, the module M4 includes the following submodules:

[0052] Module M4.1: Input the decoder output into the fully connected layer, crop the fully connected layer output, and obtain data with the same dimension as the prediction as the prediction result;

[0053] Module M4.2: Use the mean square error function as the loss function, calculate the error between the predicted value and the true value and perform backpropagation to train the network.

[0054] Preferably, the module M1 includes using CFR data as a signal for a channel state prediction task and using CSI data to calculate time coding.

[0055] Preferably, the module M1.2 includes the following submodules:

[0056] Module M1.2.1: Calculate the differential L1 and L2 norms of the CSI signal CSI∈C t×n , the calculated L1 and L2 norms CSIL∈R t :

[0057]

[0058]

[0059] Where t is the length of the time series, n is the frequency, abs(·) refers to the amplitude of the complex number, and i and j are array position indices;

[0060] Module M1.2.2: The rate of change η of the time series position t relative to the position p t,p for:

[0061]

[0062] Module M1.2.3: Using the rate of change η t,p And the selected threshold η0 is calculated to obtain the TC time code:

[0063]

[0064] Among them, i and j are array position indexes;

[0065] Module M1.2.4: Use a linear layer to map the temporal code TC to the same dimension as the embedded CFR feature. After mapping, TC∈R t×d ;

[0066] Among them, d is the dimension after mapping.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] 1. The present invention can introduce the timing code calculated by CSI during position coding in the Transformer model, so that the model can obtain more channel stability information and improve the accuracy of the model's prediction of future channel states.

[0069] 2. The present invention uses CSI calculation timing coding, and the data used is easy to obtain, the principle is simple, and it is easy to calculate. It can be applied to various Transformer-type models without changing the model structure and increasing the computational overhead, thereby enhancing the effect of signal state prediction tasks. It has broad application prospects and high practicality.

[0070] 3. The present invention incorporates time information into the model by calculating the timing characteristics of the channel state information and adding coding, thereby taking into account the relative stability time of the channel at future moments, solving the problem of weak timing information embedding ability in Transformer-type models, and improving the accuracy of channel state prediction.

[0071] Other beneficial effects of the present invention will be explained through the introduction of specific technical features and technical solutions in the specific implementation methods. Those skilled in the art should be able to understand the beneficial technical effects brought about by the introduction of these technical features and technical solutions. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Other features, objects and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments with reference to the following drawings:

[0073] Figure 1 Flow chart of the method of the present invention.

[0074] Figure 2 Schematic diagram of the channel state information prediction task in the present invention.

[0075] Figure 3 This is a schematic diagram of the network structure of the standard Transformer model in the present invention.

[0076] Figure 4 Schematic diagram of the network structure of the time-encoded Transformer model proposed in this invention.

[0077] Figure 5 This is a flow chart of the time coding calculation proposed by the present invention.

[0078] Figure 6 This is a schematic diagram of the Informer network structure for time coding proposed in this invention. DETAILED DESCRIPTION

[0079] The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the scope of the present invention. These all fall within the scope of protection of the present invention.

[0080] The method proposed in the present invention is applicable to Transformer-type models. When position coding is added after the embedding layer, additional temporal information coding is added to help the model obtain more channel state information.

[0081] The basic principle of the present invention is:

[0082] In channel state information prediction, the channel frequency response (CFR) is used as the prediction target. CFR can describe the impact of the channel on signal propagation from the perspective of amplitude-frequency characteristics and phase-frequency characteristics. With the help of the channel frequency response of a specific signal, the propagation characteristics of different multipath paths can be calculated, from which information useful for positioning and perception can be analyzed.

[0083] At the same time, the corresponding channel state information (CSI) is also required to calculate the proposed time coding. This coding uses the rate of change of the CSI sequence to describe the change in relative coherence time. The relative coherence time refers to the time required for the channel state information (CSI) value to change under a certain threshold. This integrates richer time information and is suitable for models that provide local temporal context information for the CSI sequence, such as the Transformer model.

[0084] Reference Figure 1 As shown, a coding method for a Transformer-type model for channel state information prediction includes:

[0085] Step 1: Obtain channel status time series data, including CFR and CSI;

[0086] Among them, CFR is the feature predicted by the model, and CSI is the feature of the calculated temporal encoding.

[0087] Step 2: Process the CFR, split the real part and the imaginary part and splice them into a real matrix;

[0088] Step 3: Embed the CFR through the convolutional layer and add the position encoding of the Transformer model and the time encoding proposed by this invention;

[0089] Step 4: Input the embedded and encoded data into the encoder to extract features;

[0090] Step 5: Embed the CFR data to be predicted and add the position encoding of the Transformer model without adding the time encoding. Input the encoder output and the CFR feature data to be predicted into the decoder for decoding;

[0091] Step 6: Input the encoder output into the fully connected layer and crop the fully connected layer output to obtain data with the same dimension as the prediction as the prediction result;

[0092] Step 7: Use the Mean Squared Error (MSE) as the loss function, calculate the error between the predicted value and the true value and perform backpropagation to train the network.

[0093] Step 8: Use the CFR and CSI obtained in the actual scenario as input to the model to predict the channel frequency response information for a period of time in the future.

[0094] The CFR at time t can be expressed as h t ∈C m , m represents the number of subcarriers, then the prediction task in the present invention can be expressed as follows:

[0095] (h t+1 ,h t+2 ,……,h t+l )=F(h t-r+1 ,h t-r+2 ,……,h t );

[0096] Where F(·) represents a prediction function or predictor, r is the window length of historical data used for prediction, and l is the time length to be predicted, that is, the CFR of the known length r is used to predict the CFR of the future length l.

[0097] Reference Figure 2 As shown in the figure, the encoding method proposed in this invention is mainly applied to the Transformer model. The Transformer model mainly consists of a decoder and an encoder (Encoder-Decoder) structure. The encoder is used to encode the input sequence into a context representation, and the decoder is used to generate a target sequence using the context representation. The main structures of the encoder and decoder are as follows:

[0098] 1. The encoder consists of multiple identical layers, each of which contains two sublayers: a multi-head self-attention layer and a feed-forward neural network layer. Each sublayer is followed by a residual connection and a layer normalization operation.

[0099] The multi-head self-attention layer process is as follows. Given an input sequence X:

[0100] 1. Compute multiple attention heads in parallel;

[0101] For each head t, the three matrices of the attention mechanism are calculated separately: Query value Q t 、Key value K t and Value V t :

[0102]

[0103] in, and is the weight matrix.

[0104] 2. Calculate attention weights and weighted sums independently;

[0105] For each head t, calculate the attention weight and weighted sum separately:

[0106]

[0107]

[0108] Among them, d is the dimension of Query and Key, j is the one-dimensional index of the matrix, Q, K and V are the Query, Key and Value matrices of the attention mechanism respectively, SelfAttention t is the self-attention output of the t-th head.

[0109] 3. Splice the output of multiple heads;

[0110] Concatenate the outputs of all heads together and perform a linear transformation:

[0111] MultiHeadSelfAttention(X)=Concat(SelfAttention1,SelfAttention2,…,SelfAttention n )W O ;

[0112] Among them, W O is the learned weight matrix, Concat(·) is the concatenation function, SelfAttention n is the self-attention output matrix of the n-th head.

[0113] Second, the decoder consists of multiple identical layers, each of which also contains three sublayers: a multi-head mask self-attention layer, an encoder-decoder attention layer, and a feed-forward neural network layer. Similarly, each sublayer is followed by a residual connection and normalization operation.

[0114] The difference from the encoder is that the multi-head masked self-attention layer is similar to the self-attention layer in the encoder, but with the addition of a masking mechanism, that is, shielding the information of future positions in the attention calculation. After adding the mask, the attention mechanism formula is updated to:

[0115]

[0116] Where mask is a vector whose value is -∞ or 0, which is used to set the attention after the current position to 0, d is the dimension of Query and Key, Q and K are the Query and Key matrices of the attention mechanism, and softmax(·) is the activation function.

[0117] Encoder-Decoder Attention: This layer is used to use the context information output by the encoder and the previously generated part of the target sequence information to help generate the next word in the process of generating the sequence. The difference from the multi-head self-attention is that it uses the encoder output to map the key and value values. After using the encoder-decoder attention, for the tth attention head, the query value Q t 、Key value K t and Value V t The calculation formula is updated to:

[0118]

[0119] in, and is the weight matrix, Y is the input sequence of the decoder, and Z is the output sequence of the encoder.

[0120] The process of the Transformer model is as follows:

[0121] 1. Given the encoder input sequence X, embed it into a high-dimensional space and add position encoding to obtain position information. The calculation formula of position encoding PE is:

[0122]

[0123] Where d is the dimension of Query and Key, i and j are array position indexes;

[0124] 2. Input to the encoder, and extract the feature Z through multiple sub-layers consisting of multi-head attention layers and feedforward neural network layers;

[0125] 3. Given the encoder input Y, embed it and add position encoding.

[0126] 4. The feature Z extracted by the encoder and the input Y of the decoder are fed into the decoder, and after passing through multiple layers consisting of a multi-head masked attention layer, an encoder-decoder attention layer, and a feedforward neural network layer, the output is finally obtained.

[0127] Reference Figure 3 As shown in the figure, the proposed temporal encoding is applied to the encoder input, and is added to the embedded features along with the positional encoding, using the temporal encoding Transformer model structure. Therefore, the present invention is applicable not only to Transformer models, but also to various Transformer-like models, such as Informer.

[0128] The present invention can introduce the timing code calculated by CSI during position coding in the Transformer model, so that the model can obtain more channel stability information and improve the accuracy of the model in predicting future channel states. The use of CSI to calculate the timing code makes the data easy to obtain, the principle is simple, and the calculation is easy. It can be applied to various Transformer models without changing the model structure and increasing the computational overhead, thereby enhancing the effect of signal state prediction tasks. It has broad application prospects and high practicality.

[0129] The above is a basic embodiment of the present invention. The technical solution of the present invention is further described below through a preferred embodiment.

[0130] Example 1

[0131] The specific implementation method is further explained using the Informer model as an example:

[0132] Step 1: Obtain channel state timing data, including CSI and CFR, CSR∈C t×m (t is the length of the time series, m is the number of subcarriers) is the feature predicted by the model, CSI∈C t×n (t is the length of the time series, n is the frequency) is the feature for calculating the temporal coding;

[0133] Step 2: Process the CFR and split the real part and the imaginary part Use the concat(·) function to concatenate into a real matrix CSR∈C t×2m , the splicing method is as follows:

[0134]

[0135] Step 3: Embed the CFR through a one-dimensional convolutional layer and add encoding;

[0136] Step 3.1: Embed CFR using a one-dimensional convolutional layer to map CFR to a high-dimensional space. After embedding, CSR∈R t ×d (d is the dimension after mapping);

[0137] Step 3.2: Use CSI to calculate the temporal code and the original Transformer position code CSR∈R t×d Added to the embedded features, the specific time encoding calculation method is as follows:

[0138] 1. Calculate the differential L1 and L2 norms of CSI, CSI is a complex matrix, CSR∈C t×n(t is the length of the time series, n is the frequency), the calculated L1 and L2 norms CSIL∈R t , the formula is as follows, where abs(·) refers to the amplitude of the complex number, and i and j are array position indices:

[0139]

[0140]

[0141] 2. Define the rate of change η t,p , the rate of change of the time series position t relative to the position p is defined as follows:

[0142]

[0143] 3. According to the rate of change η t,p The TC time code is calculated by the selected threshold η0. The calculation method is to compare η from large to small for each time point i and all time points j less than i. i,j Is the size of greater than the threshold η0. If there is a time point j such that η i,j <η0, then the time code value of position i is ij; if there is no position j less than i, such that η i,j <η0, the time code of position i is 0. In addition, in order to ensure that the final matrix length t remains unchanged, when i=0, the TC time code is 0 to fill the value of the first position. The final time code TC∈R t .

[0144]

[0145] Among them, i and j are array position indexes.

[0146] 4. Use the linear layer to map the time code TC to the same dimension as the feature. After mapping, TC∈R t×d .

[0147] 5. Calculate the original Transformer position encoding PE∈R t×d , the calculation formula is as follows:

[0148]

[0149] Among them, i and j are array position indexes.

[0150] 6. Add the time code TC, position code PE and embedded CFR signal to get the final enc_input data, the final enc_input∈R t×d , the formula is as follows:

[0151] enc_input=CFR+PE+TC;

[0152] Step 4: Input the enc_input data into the Informer encoder to extract features;

[0153] Step 4.1: The features input to the encoder will pass through a module consisting of a multi-head ProbSparse self-attention layer and a one-dimensional convolutional layer to learn weights, and then pass through a maximization layer to reduce the dimension;

[0154] Step 4.2: Repeat step 4.1 twice and finally get the encoder output enc_out∈R l×d (l is the feature length output by the encoder, and d is the feature dimension).

[0155] Step 5: Input the output of the encoder and the data to be predicted into the decoder for decoding;

[0156] Step 5.1: Predict the CFR of a given data set k ∈R k×m (k is the known sequence length) and a section of zero data CFR that needs to be predicted u ∈0 u×m (u is the length of the sequence to be predicted) spliced ​​together, this signal is also embedded in step 3, and only the position code is added, and finally the decoder input dec_input∈R is obtained (k+u)×d ;

[0157] Step 5.2: The embedded and encoded prediction data will be input into the multi-head ProbSparse self-attention layer, and the output intermediate result out∈R (k+u)×d With the encoder output enc_out∈R l×d Input them together into the cross attention layer, and finally get the encoder output dec_out∈R (k+u)×d ;

[0158] Step 6: Input the output of the encoder into the fully connected layer and crop the output of the fully connected layer to obtain the data with the same prediction dimension as the prediction result pred∈R u×m ;

[0159] Step 7: Use the Mean Squared Error (MSE) as the loss function, calculate the error between the predicted value and the true value and perform backpropagation to train the network.

[0160] Step 8: Use the CFR and CSI obtained in the actual scenario as input to the model to predict the channel frequency response information for a period of time in the future.

[0161] The present invention incorporates time information into the model by calculating the timing characteristics of channel state information and adding coding, thereby taking into account the relative stability time of the channel at future moments, solving the problem of weak timing information embedding ability in Transformer-type models, and improving the accuracy of channel state prediction.

[0162] The present invention also provides a coding system for a Transformer-type model for channel state information prediction. The coding system for the Transformer-type model for channel state information prediction can be implemented by executing the process steps of the coding method for the Transformer-type model for channel state information prediction. That is, those skilled in the art can understand the coding method for the Transformer-type model for channel state information prediction as a preferred implementation of the coding system for the Transformer-type model for channel state information prediction.

[0163] Specifically, a coding system for a Transformer-type model for channel state information prediction includes:

[0164] Module M1: Acquire and process channel state timing data;

[0165] The time series data includes CSI data and CFR data; the processing includes embedding and encoding;

[0166] Module M2: Inputs the processed data into the encoder to extract features;

[0167] Module M3: Embeds the CFR data to be predicted and adds the position encoding of the Transformer model; inputs the relevant parameters into the decoder for decoding;

[0168] The parameters include the output of the encoder and the CFR feature data to be predicted;

[0169] Module M4: processes the decoder output and predicts channel state information;

[0170] Module M5: Uses the CFR data and CSI data obtained in actual scenarios as input to the model to predict channel frequency response information.

[0171] The module M1 includes the following submodules:

[0172] Module M1.1: Split the CFR data into real and imaginary parts and concatenate them into a real matrix:

[0173]

[0174] in, is the real part of the CFR data, is the imaginary part of the CFR data, concat(·) is the concatenation function;

[0175] Module M1.2: Calculate time code;

[0176] Module M1.3: Embed the CFR data through the convolutional layer, add time encoding and position encoding of the Transformer model.

[0177] The module M4 includes the following submodules:

[0178] Module M4.1: Input the decoder output into the fully connected layer, crop the fully connected layer output, and obtain data with the same dimension as the prediction as the prediction result;

[0179] Module M4.2: Use the mean square error function as the loss function, calculate the error between the predicted value and the true value and perform backpropagation to train the network.

[0180] The module M1 includes using CFR data as a signal for a channel state prediction task and using CSI data to calculate a time code.

[0181] The module M1.2 includes the following submodules:

[0182] Module M1.2.1: Calculate the differential L1 and L2 norms of the CSI signal CSI∈C t×n , the calculated L1 and L2 norms CSIL∈R t :

[0183]

[0184]

[0185] Where t is the length of the time series, n is the frequency, abs(·) refers to the amplitude of the complex number, and i and j are array position indices;

[0186] Module M1.2.2: The rate of change η of the time series position t relative to the position p t,p for:

[0187]

[0188] Module M1.2.3: Using the rate of change η t,p And the selected threshold η0 is calculated to obtain the TC time code:

[0189]

[0190] Among them, i and j are array position indexes;

[0191] Module M1.2.4: Use a linear layer to map the temporal code TC to the same dimension as the embedded CFR feature. After mapping, TC∈R t×d ;

[0192] Among them, d is the dimension after mapping.

[0193] Those skilled in the art will appreciate that, in addition to implementing the system and its various devices, modules, and units provided by the present invention in purely computer-readable program code, it is entirely possible to implement the same functions of the system and its various devices, modules, and units provided by the present invention in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system and its various devices, modules, and units provided by the present invention can be considered a hardware component, and the devices, modules, and units included therein for implementing various functions can also be considered as structures within the hardware component; the devices, modules, and units for implementing various functions can also be considered as both software modules implementing the method and structures within the hardware component.

[0194] In the description of this application, it should be understood that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", etc., indicating the orientation or position relationship, are based on the orientation or position relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on this application.

[0195] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. The embodiments of this application and the features in the embodiments may be combined with each other in any manner unless there is a conflict.

Claims

1. A coding method for a Transformer-type model for channel state information prediction, characterized in that: include: Step S1: Acquire and process channel state time series data; The time series data includes CSI data and CFR data; the processing includes embedding and encoding; Step S2: Input the processed data into the encoder to extract features; Step S3: Embed the CFR data to be predicted and add the position encoding of the Transformer model; Input relevant parameters into the decoder for decoding; The parameters include the output of the encoder and the CFR feature data to be predicted; Step S4: Process the output of the decoder to predict channel state information; Step S5: Use the CFR data and CSI data obtained in the actual scenario as input to the model to predict channel frequency response information.

2. The encoding method of a Transformer-type model for channel state information prediction according to claim 1, characterized in that: The step S1 includes the following sub-steps: Step S1.1: Split the CFR data into real and imaginary parts and concatenate them into a real matrix: in, is the real part of the CFR data, is the imaginary part of the CFR data, concat(·) is the concatenation function; Step S1.2: Calculate and obtain the time code; Step S1.3: The CFR data is embedded through the convolutional layer, and the time encoding and position encoding of the Transformer model are added.

3. The encoding method of a Transformer-type model for channel state information prediction according to claim 1, characterized in that The step S4 includes the following sub-steps: Step S4.1: Input the decoder output into the fully connected layer, and crop the fully connected layer output to obtain data with the same dimension as the prediction as the prediction result; Step S4.2: Use the mean square error function as the loss function, calculate the error between the predicted value and the true value and perform backpropagation to train the network.

4. The encoding method of a Transformer-type model for channel state information prediction according to claim 2, characterized in that: The step S1 includes using CFR data as a signal for a channel state prediction task and using CSI data to calculate a time code.

5. The encoding method of a Transformer-type model for channel state information prediction according to claim 2, characterized in that: The step S1.2 includes the following sub-steps: Step S1.2.1: Calculate the differential L1 and L2 norms of the CSI signal CSI∈C t×n , the calculated L1 and L2 norms CSIL∈R t : Where t is the length of the time series, n is the frequency, abs(·) refers to the amplitude of the complex number, and i and j are array position indices; Step S1.2.2: The rate of change η of the time series position t relative to the position p t,p for: Step S1.2.3: Use the rate of change η t,p And the selected threshold η0 is calculated to obtain the TC time code: Among them, i and j are array position indexes; Step S1.2.4: Use a linear layer to map the temporal code TC to the same dimension as the embedded CFR feature. After mapping, TC∈R t×d ; Among them, d is the dimension after mapping.

6. A coding system for a Transformer-type model for channel state information prediction, characterized in that: include: Module M1: Acquire and process channel state timing data; The time series data includes CSI data and CFR data; the processing includes embedding and encoding; Module M2: Inputs the processed data into the encoder to extract features; Module M3: Embeds the CFR data to be predicted and adds the position encoding of the Transformer model; inputs the relevant parameters into the decoder for decoding; The parameters include the output of the encoder and the CFR feature data to be predicted; Module M4: processes the decoder output and predicts channel state information; Module M5: Uses the CFR data and CSI data obtained in actual scenarios as input to the model to predict channel frequency response information.

7. The coding system of a Transformer-type model for channel state information prediction according to claim 6, characterized in that: The module M1 includes the following submodules: Module M1.1: Split the CFR data into real and imaginary parts and concatenate them into a real matrix: in, is the real part of the CFR data, is the imaginary part of the CFR data, concat(·) is the concatenation function; Module M1.2: Calculate time code; Module M1.3: Embed the CFR data through the convolutional layer, add time encoding and position encoding of the Transformer model.

8. The coding system of a Transformer-type model for channel state information prediction according to claim 6, characterized in that: The module M4 includes the following submodules: Module M4.1: Input the decoder output into the fully connected layer, crop the fully connected layer output, and obtain data with the same dimension as the prediction as the prediction result; Module M4.2: Use the mean square error function as the loss function, calculate the error between the predicted value and the true value and perform backpropagation to train the network.

9. The coding system of a Transformer-type model for channel state information prediction according to claim 7, characterized in that: The module M1 includes using CFR data as a signal for a channel state prediction task and using CSI data to calculate a time code.

10. The coding system of a Transformer-type model for channel state information prediction according to claim 7, characterized in that: The module M1.2 includes the following submodules: Module M1.2.1: Calculate the differential L1 and L2 norms of the CSI signal CSI∈C t×n , the calculated L1 and L2 norms CSIL∈R t : Where t is the length of the time series, n is the frequency, abs(·) refers to the amplitude of the complex number, and i and j are array position indices; Module M1.2.2: The rate of change η of the time series position t relative to the position p t,p for: Module M1.2.3: Using the rate of change η t,p And the selected threshold η0 is calculated to obtain the TC time code: Among them, i and j are array position indexes; Module M1.2.4: Use a linear layer to map the temporal code TC to the same dimension as the embedded CFR feature. After mapping, TC∈R t×d ; Among them, d is the dimension after mapping.

Images